Abstract
One of the fundamental challenges in the field of computer vision is segmenting what the user-preferred region of interests (ROIs) of a given image. Segmenting the ROIs for an image is not a trivial task, especially for images that contain multiple objects and cluttered backgrounds. Interactive image segmentation, or image segmentation with a human in the loop, can make the region of interest (ROI) more clearly defined for obtaining accurate segmentation. However, most of the existing interactive image segmentation algorithms rely on the user to provide accurate annotations as the guidance, and the segmentation results usually greatly depend on the qualities of the user-provided annotations. This thesis aims to introduce three efficient algorithms to make the non-expert users can obtain high-quality image segmentations with machine assistance. We develop three efficient algorithms to address the problem of interactive image segmentation. The first algorithm aims to improve the effectiveness and efficiency of interactive 1-bit user feedback image segmentation. The most interesting property is that the responsibility to define the annotation locations is transferred from the user to the machine. Then, we introduce a seed proposals approach in an interface manner for addressing the interactive segmentation task on a small touchscreen device. The interface aims to address the problem that the precise annotations are difficult to provide while annotating on the small display device. Finally, we present a new segmentation algorithm that leverages the latent photographic information available at the moment of taking pictures. A learning-based segmentation approach is proposed to collect available cues while taking pictures for carrying out the interactive segmentation in a common tap-and-shoot photographing process. The performances of our algorithms are evaluated on several publicly available datasets. The evaluation results show that our algorithms achieve high segmentation accuracy, with short response time and fewer user feedback. The evidence that our algorithms can provide the superior solutions to the interactive image segmentation problem is clear.